On 03/06/20 08:50, Vincent Guittot wrote:

On Wed, 3 Jun 2020 at 09:34, Benjamin GAIGNARD [off-list ref] wrote:

quoted

On 6/2/20 3:35 PM, Valentin Schneider wrote:

quoted

AFAIA streaming_start() is not necessarily executing on the same CPU as the
one that will handle the interrupt. I was thinking you could use the IRQ's
effective affinity as a hint of which CPU(s) to boost, i.e. something like:

---
     struct cpumask_var_t visited;
     struct irq_data *d = irq_get_irq_data(irq);

     err = alloc_cpumask_var(visited, GFP_KERNEL);
     /* ... */
     for_each_cpu(cpu, irq_data_get_effective_affinity_mask(d)) {
             /* check if not already spanned */
             if (cpumask_test_cpu(cpu, visited))
                     continue;

             policy = cpufreq_cpu_get(cpu);
             cpumask_or(visited, visited, policy->cpus);
             /* do the boost for that policy here */
             /* ... */
             cpufreq_cpu_put(policy);
     }
---

That of course falls apart when hotplug gets involved, and the effective
affinity changes... There's irq_set_affinity_notifier() out there, but it
seems it's only about the affinity, not the effective_affinity, I'm not
sure how valid it would be to query the effective_affinity in that
notifier.

quoted

If I wait to be in the irq it will be too late so I think I will do a
loop over all possible CPUs
before start the streaming to change the policies.

Yes, that's what I was thinking as well.

Can't you use irq_get_affinity_mask  and loop over it ?

In the end that's the only usable option, I think.

I was looking at alternatives because on arm64 (and AFAICT that applies
to arm too; see irq-gic.c::gic_set_affinity()) the affinity mask spans
all CPUs by default, while the effective affinity mask spans only the
CPU that will actually handle the IRQ (+ where its thread should run).

That said, using the effective mask that way does feel like an
implementation leak. Sadly I couldn't find any better way to minimize
the number of boosted frequency domains.

Also You should better use freq_qos_add/remove_request during probe
and remove of the driver and use freq_qos_update_request in
dcmi_start/stop_streaming to set/unset your constraint.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel